Anchor Box Generation

Object detectors don't guess boxes from nothing. They start from a fixed grid of anchors — reference boxes at every location, in a spread of sizes and shapes — and only learn how to nudge them.

Around a centre point, each (scale, ratio) pair defines one anchor:

1width  = scale * sqrt(ratio)
2height = scale / sqrt(ratio)

Task: write generate_anchors(center_x, center_y, scales, aspect_ratios) returning boxes as [x1, y1, x2, y2] (top-left, bottom-right), every value rounded to 4 decimal places.

• Produce one anchor per (scale, ratio) combination, looping scales on the outside, ratios on the inside.
• The box is centred on (center_x, center_y) — half the width each side, half the height above and below.
• Empty scales returns an empty list.

That sqrt in both formulas isn't decoration. It keeps width * height = scale² for every ratio, so changing shape doesn't quietly change area — the two knobs stay independent.